0:03 All right. So, little sensational title, right? But I think that there really is a bit of a battle that we need to be thinking about when it comes to our use of AI in design and computation. And this battle, if we think about it as a battle between humans and AI, is a little silly because it’s kind of like stop hitting yourself, right? Like we’re making the AI and we are in conflict with it.

0:29 So I think to resolve this battle, it largely depends on us being a little braver than we might be right now and putting our foot down to protect things that are important for the soul of design and for humans as part of design. So at CMU I have the pleasure of leading the human plus AI design initiative which is a cross-campus initiative that draws from the college of engineering, our school of computer science through the human computer interaction institute as well as our ter school of business which is where our organizational and behavioral psychology team sits.

1:02 We also draw from CMU Pittsburgh, CMU Africa, Qatar and all of our other locations to take a global view at these challenges. Now, our work is primarily driven by two objectives. I won’t talk too specifically to the work that we do in the institute. But we start with designing novel AI agents that can support and assist design in some way. And then we field those agents and examine how they work with human designers, how they change and modify the work that those designers do.

1:35 And in the space of possibilities for this human AI collaboration and human AI interaction, we simplify it a little bit, but I think this 2x two captures a lot of the interesting possibilities for human AI collaboration and interaction in design. Along the left side, we have problem focused versus process focused. I think we’re all familiar with problem focused AI, right? It’s going to design a new part for you or provide a recommended solution for your problem.

2:06 Process focused AI might say, “Hey, Duann, you haven’t had a meeting with your planning committee for like 2 months. You should really check in with them, right?” No, no planning committee, one man show, one man and a wife, right? So, problem focused versus process focus. We can also think about reactive versus proactive, right? A reactive application is something where you click a button and get a response, right?

2:29 Something more proactive might shoot you an email or a Slack message or ask you to do some work on its behalf. And when it comes to the possibilities in human AI teaming and when it comes to the future and the direction that we’re taking, I have been thinking a lot about curling. So who knows what curling is? Okay, good. Who has no idea what curling is? Okay, so we’ll do a little little intro.

2:56 So curling is kind of like shuffle board or ponk, botchi ball, pool, except you get these big old stones and these are like 30 or 40 lbs and you slide them down the ice, right? There’s a lot of momentum involved, right? You have your main pusher who they do this like great stance when they’re doing it. And they push the rock at a very specific trajectory, sometimes with very specific rotation down the court.

3:24 It slides across this ice. And one thing I actually learned this morning is that this ice isn’t the kind of ice that we think about when we, you know, might be playing hockey, right? This is actually pretty dimpled, rough, uneven ice. And because of that, you have a couple people on your team with these brushes or brooms that sort of slide down next to the stone and they can’t touch the stone, but what they can do is rub the ice in front of it and try to make the ice a little nicer and try to adjust the trajectory of the stone in that way.

3:53 And eventually these stones reach the target. They might collide with each other and you’re trying to score points and achieve the right outcome for what you want to occur. The challenge is there’s a lot of ice between where you start and where you end up. There is a degree of uncertainty, right? Professionals have maybe a little more certainty in the way that they push down the ice, the way that they’re able to impart a trajectory, but we need to worry about the initial conditions, right?

4:27 Then along the way, we need to have some idea of the dynamics of the system, right? What the ice is like, what our broom people can do. And then all of that together is going to dictate what the most likely outcomes are. And this probably isn’t the most opaque analogy, right? You’ve probably already seen the connection to AI and the way that we think about development and deployment of technology.

4:49 So as we sit here today, we have a pulse on the initial conditions. We have some control over the initial conditions. We also need to think about the dynamics of the system though. There are going to be a lot of people this week talking about AI and the different possibilities that we have with AI. I think that some of the most important dynamics of the system though are people and what we know about people.

5:14 And then all of this dictates what outcomes we’re likely to achieve. And at the end of my talk, I’ll share a few different possible futures that I think we might be headed for. So starting off, what are our initial conditions? What do we know about them? So on a corporate level we’ve seen you know recent upticks in Gen AI and also AI adoption right none of this is too secret or too new for this audience if we extrapolate and look at GDP impact over you know the next 5 or 10 years it’s going to increase AI is going to be huge there’s going to be a lot of money involved and effort put behind this this specifically shows global GDP impact some people have started to think about okay if we unschool spool this across the world.

5:58 Where are we going to see the most impact? Where are we going to see the least impact? And this map, the more you look at it, the more frustrating it gets. But there’s a lot of good stuff here. So, what this tells us is generally we’re going to expect that the global north sees a lot of the benefit from AI. There’s going to be a lot focused in Europe, North America, China and relatively small impact from Gen AI or AI in the global south.

6:26 So relatively little in South America, Africa, Oceanania. So this is starting to point out some of the inequities that we might need to think about as we develop AI in the future. We also are thinking a lot and this is perhaps my bias, but thinking a lot about how we can bring human and AI agents together in the right way. And this is an example from something that’s near and dear to my heart, which is education.

6:48 And there are ways that we can reallocate teacher time towards more meaningful tasks, right? There’s a lot of busy work that teachers do. But if we can reallocate that time using AI and help instructors engage with students with that new time, we can achieve better outcomes. So, all of this points to the fact that we have a current trajectory that is starting to think about human AI collaboration in deep and meaningful ways.

7:17 However, we also need to be better estimating in a sense our gradient where we’re at today. And one important way that we can do that is by creating better benchmarks, better data sets and sharing those with the community. And I’d actually say this is some of the most important work that we can be doing right now. A couple of the benchmarks that we’ve put out from my lab.

7:38 So, one of these is a couple years old, Megaplow 2D. And it’s a few million paired frames. So this was intended for super resolution where you take a lowresolution simulation and then enhance it using some physics- based information. Another data set or benchmark that we’re releasing soon is open ceme. So opense seeing simulations for engineering. So, it’s a visual question answering benchmark that’s based on a couple or 200,000 different question and answer pairs based on ANIs simulations.

8:15 So, these types of benchmarks and data sets help us to better understand our current gradient where we sit in AI deployment and also to improve that gradient. However, as I’ve alluded to, the AI is only one part of the equation. You also need to be thinking about how humans are adopting this technology and measuring human progress and the human gradient in human AI teaming. And one of the pieces that my team has started to do is build out metrics for looking at AI readiness and AI adoption in industry.

8:43 This is perhaps one of the most important parts of the gradient that we can try to estimate right now because this this tells us where we should be spending our effort and where we should be focusing as we continue to develop these technologies. So that’s just a little bit of where we’re at right now. And over the course of the next couple days, you’re going to see a lot more examples of the current state of artificial intelligence as it relates to computational design.

9:08 Now I want to shift a little bit and come to the middle question. So what do we know about the dynamics of the system? Well, technology changes really fast. So I love this plot. I actually had to cut off the bottom of it. So if you follow this back, it will go back to like the earliest eras of humankind and all of the technology that happened along the way.

9:33 But it is accelerated in some ways, right? We’re in a you know an exponential takeoff scenario. Some of the work that you saw in the first keynote is looking at how we can use AI to build better AI and better solutions. So technology is changing fast and even faster. So personally I don’t think this is the best place to look for trends or to try to make predictions.

9:57 Now the other place we can look is human beings. And human beings change relatively slowly. And this is you know one example but global height hasn’t really changed appreciably over the last two dec two centuries, right? Trivial example, but humans change relatively slow, right? We’re still sort of shackled with these monkey brains that are trying to deal with a new computational age and all the information that we’re being pushed into.

10:24 So I think that there’s a lot we can learn by looking at human nature. Now I have five cognitive attributes that I think we should be thinking about as we look to the future of artificial intelligence. The first is that we satisfice readily in the real world. So satisficing is a portmanto of satisfy and suffice. And it’s this idea that in the real world, we can never really achieve perfection.

10:50 If we do, we’re going to bankrupt the company because it’s going to take way too long and we’re going to spend way too much money. So we have some bounded rationality that says, okay, these are the reasonable bounds on effort that we’re going to accept and these are the reasonable bounds on value that we’re going to accept. And why is this important for how we work with AI?

11:09 AI in some cases gives us a very ready path towards satisficing towards getting something that’s just good enough. So this is a deep part of how we think as humans that we need to be thinking about as we move forward. What are the opportunities that we’re giving people to satisfice as opposed to improve. Second, we tend to be biased in favor of automation. And this is probably one of the factors that is changing most quickly right now.

11:36 But this has been shown across a variety of different cases. We tend to prefer automated recommendations. And this has been shown in medicine, this has been shown in manufacturing. A lot of different areas. We tend to accept automated recommendations. And over time, if those tend to be pretty good, we actually start to shut down the part of our brain that’s monitoring that AI. So we start to think less critically about the recommendations that it’s giving and that opens us up for a lot of potential error and a lot of potential risk.

12:10 So this idea of automation bias is another thing that’s really important for our future in AI. Third, we neglect the value of subtractive changes. Almost always we like to put band-aids on things and we like to add to solutions to solve problems rather than subtracting away. And there are two ways that this has been studied recently. One way is these dot problems. So the question is, okay, for any of these, you can click on a square to turn it on or turn it off.

12:41 And your task is to make these symmetric as quickly as possible and as efficiently as possible. And depending on the complexity, depending on the time pressure, people will almost always prefer additive solutions to subtractive. On the other side, this is actually one of the the more interesting tasks. So, they actually gave people this little Lego setup. And you can’t really see because of the angle, but the top shelf is sort of canolvered out, right?

13:11 So, very unstable. And the task is, okay, this shelf needs to support some weight right over the head of your little Lego dude. How do you make that happen? And the obvious solution that a lot of people go for is, okay, we’re going to add a column right here and make sure that this roof is nice and strong and it’s not going to fall on our Lego guy.

13:33 There’s a subtractive solution. We could just remove this block in the back, set the roof down on a very stable surface. Something like 20% of people saw that solution. 80% added a column. So, this is deep. We almost always look for additive solutions as opposed to subtractive. And this is no comment on additive and subtractive manufacturing. Fourth, humans were almost always copy first movers. So being at CMU, has anybody read the last lecture or watched the last lecture?

14:09 Okay, so look it up on YouTube if you haven’t. In it, Randy Pouch, former CMU professor, makes this case that we always want to be the first penguin, right? So, we want to be the first mover. We want to take the risk. However, the problem with that is that people will follow you. And sometimes they don’t do it for good reasons and they don’t do it for the right reasons.

14:29 So it’s sort of imminent on us to think about if we are the early mover, are we doing it for the right reasons? If we aren’t the early mover and we’re a second mover, are we doing that for the right reasons as well? Fifth, we suffer from choice overload. And this is something that I think we have all experienced, analysis paralysis, right? And this is something that is probably going to become an increasing issue with Gen AI.

14:59 A predominant paradigm for Gen AI is to present a set of solutions which is great but it means you have to reason over those same solutions. So that’s a little bit about what I think are the important dynamics of the system, right? What we know about human nature that sort of tells us how this OD is going to evolve. The last piece I want to look at is potential futures.

15:24 So to start off, does anybody know what P Doom is? Have you heard Pd Doom? Yeah. So pdoom describes your personal probability for doom of the human race as a result of AI. So one possible future is well pd doom approaches one terrible things happen to humanity and we’re screwed, right? Not the best potential future. Depending on how you feel, you might think we’re headed directly to this or not.

15:57 Another potential future is as a species we aggressively drive pd doom to zero. We shut off all the computers and we just embrace this idealic lifestyle. H there’s something to like about that but maybe not the the best future. Sort of an intermediate future is yeah maybe we become cy we become cyborgs right? With Neurolink we’re starting to see some steps towards this. But depending on who you talk to in human computer interaction, the internet has made us cyborgs decades ago.

16:29 So this might be a future we’re headed for. Now getting into some more I’d say realistic potential futures. I think one potential future is that humans will end up doing the physical parts of design. So there’s a lot of research right now focused on fine motor manipulation for robotics. It is a huge open challenge and we as people walk around with these things attached to our bodies usually and these are worth millions of dollars to roboticists, right?

17:01 So we might if AI continues to progress in a cognitive and innovation space be relegated to a physical world because we have fine motor manipulation. Potential future five, we might end up doing the early stages of design. This is a bit of what we’re seeing already, right? A lot of design techniques, design technology supports the later stages with AI for detailed design, manufacturing, etc. And the early stages tend to be a little less supported.

17:30 So that might be a place where we decide to draw a line. Potential future number six, maybe we just end up doing the parts of design that we like the most. If we talk to my students at CMU, this is probably going to mean that they just want to play around in CAD all day. And I love playing around in CAD. But what does this future require?

17:52 This probably means that we’ve achieved some sort of universal basic income, right? We’re in a post scarcity future and we can pursue the tasks that we enjoy and that we like doing. So my preferred future potential future number seven is that humans are going to end up doing the most human parts of design. We have the we have a monopoly on empathy. Empathy is something that people have tried to automate and failed at.

18:21 And I believe deeply that our ability to relate to our fellow human beings is something that is intrinsically human. So I think that as we continue to advance the future of AI, we need to think deeply about what it means to be human and protect those aspects for ourselves and maintain the soul of design. So with that, I’d like to give a quick shout out to the amazing team that I’ve been able to work with at CMU and thank you very much. I’m looking forward to speaking with you. Unfortunately, I don’t have time for questions.